Peters Baron, Beckham Gregg T, Trout Bernhardt L
Centre Européen de Calcul Atomique Moléculaire (CECAM), Ecole Normale Supérieure, 46 Allée d'Italie, 69364 Lyon Cedex 7, France.
J Chem Phys. 2007 Jul 21;127(3):034109. doi: 10.1063/1.2748396.
This paper extends our previous work on obtaining reaction coordinates from aimless shooting and likelihood maximization. We introduce a simplified version of aimless shooting and a half-trajectory likelihood score based on the committor probability. Additionally, we analyze and compare the absolute log-likelihood score for perfect and approximate reaction coordinates. We also compare the aimless shooting and likelihood maximization approach to the earlier genetic neural network (GNN) approach of Ma and Dinner [J. Phys. Chem. B 109, 6769 (2005)]. For a fixed number of total trajectories in the GNN approach, the accuracy of the transition state ensemble decreases as the number of trajectories per committor probability estimate increases. This quantitatively demonstrates the benefit of individual committor probability realizations over committor probability estimates. Furthermore, when the least squares score of the GNN approach is applied to individual committor probability realizations, the likelihood score still provides a better approximation to the true transition state surface. Finally, the polymorph transition in terephthalic acid demonstrates that the new half-trajectory likelihood scheme estimates the transition state location more accurately than likelihood schemes based on the probability of being on a transition path.
本文扩展了我们之前关于通过无目标射击和似然最大化来获取反应坐标的工作。我们引入了无目标射击的简化版本以及基于反应概率的半轨迹似然得分。此外，我们分析并比较了完美反应坐标和近似反应坐标的绝对对数似然得分。我们还将无目标射击和似然最大化方法与Ma和Dinner [《物理化学杂志B》109, 6769 (2005)] 早期的遗传神经网络 (GNN) 方法进行了比较。在GNN方法中，对于固定的总轨迹数，每个反应概率估计的轨迹数增加时，过渡态系综的准确性会降低。这从定量上证明了单个反应概率实现相对于反应概率估计的优势。此外，当将GNN方法的最小二乘得分应用于单个反应概率实现时，似然得分仍然能更好地逼近真实的过渡态表面。最后，对苯二甲酸中的多晶型转变表明，新的半轨迹似然方案比基于处于过渡路径概率的似然方案更准确地估计了过渡态位置。
